TotalSegmentator MRI: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in MRI (2405.19492v2)

Abstract: Since the introduction of TotalSegmentator CT, there is demand for a similar robust automated MRI segmentation tool that can be applied across all MRI sequences and anatomic structures. In this retrospective study, a nnU-Net model (TotalSegmentator) was trained on MRI and CT examinations to segment 80 anatomic structures relevant for use cases such as organ volumetry, disease characterization, surgical planning and opportunistic screening. Examinations were randomly sampled from routine clinical studies to represent real-world examples. Dice scores were calculated between the predicted segmentations and expert radiologist reference standard segmentations to evaluate model performance on an internal test set, two external test sets and against two publicly available models, and TotalSegmentator CT. The model was applied to an internal dataset containing abdominal MRIs to investigate age-dependent volume changes. A total of 1143 examinations (616 MRIs, 527 CTs) (median age 61 years, IQR 50-72) were split into training (n=1088, CT and MRI) and an internal test set (n=55; only MRI), two external test sets (AMOS, n=20; CHAOS, n=20; only MRI), and an internal aging-study dataset of 8672 abdominal MRIs (median age 59 years, IQR 45-70) were included. The model showed a Dice Score of 0.839 on the internal test set and outperformed two other models (Dice Score, 0.862 versus 0.759; and 0.838 versus 0.560; p<.001 for both). The proposed open-source, easy-to-use model allows for automatic, robust segmentation of 80 structures, extending the capabilities of TotalSegmentator to MRIs of any sequence. The ready-to-use online tool is available at https://totalsegmentator.com, the model at https://github.com/wasserth/TotalSegmentator, and the dataset at https://zenodo.org/records/14710732.

• The paper introduces a sequence-independent MRI segmentation model that automates the delineation of 59 anatomical structures, reducing manual labor.
• The methodology leverages a multi-modal training approach with MRI and CT scans using an iterative learning framework based on nnU-Net, achieving a Dice score of 0.824.
• The results demonstrate improved clinical workflow efficiency and robustness, outperforming comparable segmentation models on both MRI and CT images.

TotalSegmentator MRI: Sequence-Independent Segmentation of 59 Anatomical Structures in MR Images

The paper "TotalSegmentator MRI: Sequence-Independent Segmentation of 59 Anatomical Structures in MR Images" addresses a critical gap in the current state of automated medical image segmentation by extending the functionalities of the existing TotalSegmentator framework to Magnetic Resonance Imaging (MRI). This research seeks to alleviate the labor-intensive and error-prone process of manual MRI segmentation, enhancing the workflow in clinical and research environments.

Motivation and Context

Magnetic Resonance Imaging is paramount in medical diagnostics for its detailed, ionizing radiation-free imaging of the human body. However, the manual segmentation process is cumbersome and inconsistent due to variations in interrater reliability. Existing automated segmentation tools like nnU-Net have made strides, particularly in CT image segmentation, but the diverse nature of MRI protocols injects additional complexities that these tools struggle with.

Data and Methods

The dataset for model training included 298 MRI scans and 227 CT scans, ensuring a rich variety of imaging parameters and anatomical diversity. This methodology leverages the robustness of the nnU-Net framework, known for its adaptive architectural and preprocessing configurations. By employing an iterative learning approach, the research team generated a comprehensive ground truth for 59 anatomical structures across various MRI sequences.

Experimental Results

The model's performance, evaluated using the Dice similarity coefficient (Dice), demonstrated robust segmentation capabilities. On the MRI test set, encompassing intricate clinical data with major pathologies, the model achieved a Dice score of 0.824 [CI: 0.801, 0.842]. This performance significantly surpassed that of other publicly available models like MRSegmentator and AMOS, which scored 0.762 and 0.542 respectively (p<0.001 for both comparisons).

Moreover, when tested on CT images from the original TotalSegmentator dataset, the model nearly equaled the efficiency of the original TotalSegmentator (Dice score 0.960 versus 0.970; p<0.001), underscoring its cross-modality robustness. Despite some observed failure cases due to lower image quality in MRI, especially in highly anisotropic images, the model maintained a credible level of accuracy and reliability.

Implications and Future Directions

The practical implications of these results are manifold. Clinically, the TotalSegmentator MRI can considerably reduce radiologists' workload and enhance diagnostic accuracy through consistent and rapid segmentation. The model's ability to handle a broad spectrum of MRI sequences without sequence-specific tuning also elevates its adaptability in real-world scenarios.

Theoretically, these findings highlight the synergy of multi-modal training datasets (MRI and CT) in augmenting segmentation performance. The observed benefits of integrating CT scans into the training process suggest a promising direction for further improving model robustness across different imaging modalities.

Future research could expand this work by incorporating additional anatomical structures, refining ground truth annotations, and enlarging the training dataset to encompass even more diverse pathologies and imaging variations. Moreover, continued investigation into optimizing memory and computational efficiency will be crucial for widespread clinical integration.

Conclusion

The paper successfully extends the TotalSegmentator framework to MRI images, providing a versatile and high-performing tool for the automatic segmentation of 59 anatomical structures. This open-source model, backed by publicly available training data and resources, stands out for its ease of use, clinical relevance, and robust performance, setting a new benchmark for automated MRI image segmentation.

References

The full list of references used in the paper can be found in the original paper. Key references include works on nnU-Net by Isensee et al., methodologies for MRI segmentation, and various clinical data collected from international repositories such as Imaging Data Commons and The Cancer Imaging Archive.